A Biologically-Inspired Framework for Contour Detection Using Superpixel-Based Candidates and Hierarchical Visual Cues

Xiao Sun, Ke Shang, Delie Ming, Jinwen Tian, Jiayi MaSensors 2015 · OpenReview

TL;DR

This paper introduces a novel biologically-inspired framework for contour detection that leverages superpixels to generate contour candidates and then weights these candidates using hierarchical visual cues. This approach aims to improve contour detection accuracy and efficiency by moving beyond pixel-level processing and incorporating principles from human visual perception.

What Problem Does This Paper Address?

Contour detection is a fundamental problem in computer vision, crucial for tasks like object recognition and segmentation. Traditional methods often rely on pixel-level analysis, which can be computationally expensive and prone to generating noisy or discontinuous contours, especially in complex scenes with clutter or textures. While biologically-inspired models have shown promise, they often still operate at the pixel level. The challenge is to develop a method that can efficiently and accurately detect meaningful contours by mimicking aspects of the human visual system's ability to integrate information and group features into coherent structures.

Methods

The proposed framework, called Biologically-Inspired Candidate Weighting (BICW), tackles contour detection in three main stages:

  1. Superpixel-Based Contour Candidate Generation: Instead of processing individual pixels, the authors first segment the image into superpixels using a modified Simple Linear Iterative Clustering (SLIC) algorithm. The boundaries of these superpixels are then treated as initial contour candidates. This approach reduces the search space and generates more coherent, closed contour segments. The modification to SLIC involves running it with multiple initial cluster centers to ensure that stable edges (likely to be true contours) are captured more robustly across different initializations.
  2. Hierarchical Visual Cue Extraction and Weighting: The generated contour candidates are then weighted based on a hierarchy of visual cues, inspired by the human visual system:
  3. Low-Level Cues: These cues assess the intrinsic properties of a contour candidate. The authors extract luminance and chromatic contrast energy from the superpixels adjacent to the candidate, using the CIELAB color space.
  4. Mid-Level Cues (Gestalt Principles): These cues evaluate the grouping constraint of contour candidates, mimicking how humans group features. They incorporate principles like:
  5. Curvature Constancy: Evaluating how smoothly a contour bends.
  6. Continuation: Assessing if neighboring contour candidates align in a linear or co-circular path.
  7. Similarity: Measuring the contrast difference between adjacent contour candidates.
  8. Multi-Scale Integration: The entire process of candidate generation and weighting is performed at multiple scales (three scales in this work). The final contour map is obtained by combining the weighted candidates from each scale, leveraging the idea that salient contours are likely to be consistent across different resolutions.

Datasets

The proposed method was evaluated on the Berkeley Segmentation Dataset (BSDS), specifically the BSDS300 and BSDS500 datasets. These datasets are standard benchmarks for contour detection and provide ground truth labels from multiple human subjects. The authors used the training set of BSDS500 for parameter tuning and the test set for final evaluation.

Key Insights and Ideas

  • Superpixels as Contour Primitives: The core idea of using superpixel boundaries as initial contour candidates is a significant departure from pixel-based methods. This effectively pre-groups pixels into meaningful regions, reducing noise and computational load.
  • Biologically-Inspired Hierarchical Cues: The framework's strength lies in its multi-level approach to weighting candidates, drawing parallels to how the human visual system processes information from low-level features to higher-level grouping principles (Gestalt laws).
  • Multi-Scale Consistency Boosts Salience: Integrating information from multiple scales amplifies the weights of contours that are consistently detected across different resolutions, helping to distinguish strong contours from clutter.
  • Trade-off Between Performance and Efficiency: The superpixel approach, combined with hierarchical cues, offers a promising balance between achieving high detection accuracy and maintaining computational efficiency, especially compared to some complex global methods.

Key Results

  • The modified SLIC algorithm for superpixel generation achieved a high boundary recall of 0.98 at 4000 superpixels, demonstrating its effectiveness in capturing ground truth edges.
  • On the BSDS500 dataset, the proposed BICW framework achieved an optimal F-measure of 0.68 on the training set.
  • When compared to state-of-the-art methods on BSDS300 and BSDS500, the BICW framework achieved competitive results, outperforming methods like Pb, CO, and MCI. For instance, on BSDS500, it achieved an F-score of 0.68, comparable to the Pb method (0.67) but with significantly less computational cost.
  • The method demonstrated a good trade-off between performance and efficiency, with a computation time of 13 seconds per contour map on a standard Intel Core 2 processor, significantly faster than methods like gPb (191 seconds).
  • Qualitative results showed that the method effectively suppresses "weak" boundary labels (clutter or ornaments) and is consistent with strong boundaries in the consensus labels.

Limitations and Future Work

  • Potential for "Weak" Boundary Suppression: The authors note that their method tends to suppress "weak" boundary labels, which might be undesirable if these labels are considered meaningful in certain contexts. This is also highlighted as a potential pitfall of the BSDS benchmark itself.
  • Parallel Computing Suitability: The authors suggest that their method is highly suitable for parallel computing, indicating this as a future direction for further optimization and speedup.
  • Integration of Other Visual Features: The framework is designed to be extensible, allowing for the integration of other visual features in the future.
  • Texture Information: While the method performs well, it does not explicitly use texture information, unlike some other state-of-the-art methods like Pb. Future work could explore incorporating texture cues.

Summary updated

Abstract

Contour detection has been extensively investigated as a fundamental problem in computer vision. In this study, a biologically-inspired candidate weighting framework is proposed for the challenging task of detecting meaningful contours. In contrast to previous models that detect contours from pixels, a modified superpixel generation processing is proposed to generate a contour candidate set and then weigh the candidates by extracting hierarchical visual cues. We extract the low-level visual local cues to weigh the contour intrinsic property and mid-level visual cues on the basis of Gestalt principles for weighting the contour grouping constraint. Experimental results tested on the BSDS benchmark show that the proposed framework exhibits promising performances to capture meaningful contours in complex scenes.
BibTeX
@inproceedings{sun2015biologically,
  author = {Xiao Sun and Ke Shang and Delie Ming and Jinwen Tian and Jiayi Ma},
  title = {{A Biologically-Inspired Framework for Contour Detection Using Superpixel-Based Candidates and Hierarchical Visual Cues}},
  booktitle = {Sensors},
  year = {2015},
  url = {https://openreview.net/forum?id=SMcTDdYdmF},
}
Figure